Comparison of methods for interpolating soil properties using limited data

Soil model prediction area, Omo National Park, Ethiopia. Click on image to enlarge.

Spatial interpolation of biological and biophysical variables is of central interest to many landscape ecologists. A variety of interpolation methods are currently in use and there is debate about which produces the most reliable predictions. Most of the debate centers around the influence of data characteristics and sample size on accuracy, with most comparisons based on numerous fine-spaced samples. In this study we evaluated the utility, accuracy, and effectiveness of ordinary kriging, inverse-distance weighting, and thin-plate smoothing splines with tensions using limited, coarse-spaced soil samples from the Omo Basin, Ethiopia (see Fig.1). We also endeavored to determine whether and how these method differed in accuracy and effectiveness.

Accuracy was based on the mean absolute and mean square error measures, and effectiveness was based on the goodness-of-prediction measure. Ordinary kriging and inverse-distance weighting performed the best, with each deriving similar predictions and encountering similar problems when predicting for specific locations. Thin-plate smoothing splines with tensions was the least reliable method, however, because it both smoothed the data and produced estimates for the edges of the grid that were of a greater magnitude of error. Outliers were the only characteristics that appeared important. Their importance varied however, depending on whether they augmented or diminished spatial dependence.

Using limited data in this exercise not only led to satisfying results, it also proved reliable for deriving landscape-scale predictions. These findings advocate the use of interpolation methods in situations where data are coarse-spaced and limited, constraints under which many landscape ecologists operate (see Fig.2).

Participants

Contact

• Niklaus Zimmermann